[Comp-neuro] useful models and the scientific method
dbeeman at dogstar.colorado.edu
Sat Aug 16 06:24:44 CEST 2008
As a reformed physicist turned realistic neural modeling zealot,
thought I'd throw my two cents into this discussion.
Physics has been enormously successful in giving simple explanations
with simple equations for much that surrounds us. We can explain
a lot with Newton's laws and the three laws of thermodynamics without
needing to know anything about molecules or quantum mechanics.
But this success has made physicists feel that all problems, including
"understanding the brain" can be solved with this macroscopic approach.
When I taught freshman physics I got the inevitable crank letters from
people who claimed they had a sure-fire idea for a perpetual motion
machine. The third law made it easy to dismiss these if I didn't feel
like looking for the specific flaw in the proposed solution. Yet, I never
felt that I really understood thermodynamics until I taught statistical
thermodynamics, deriving thermodynamics from the statistical mechanics of
As I began modeling liquids, glasses, and other disordered systems, the
top-down macroscopic approach became less useful. The only way to
understand an amorphous semiconductor is to make good a structural model
and good models for the ways the atoms interact. Then the model can help
one understand its properties. In fact a lot of interesting current
topics in physics are dealt with in this way. If you are going to
understand atmospheric turbulence, all the equations of classical
hydrodyamics aren't going to help you. You have to have good models,
study the behavior of the models, and try to develop some intuition for
what is happening.
So, the situation isn't much different in computational neuroscience. I
think that studies of huge integrate-and-fire networks or connectionist
models of memory are unlikely to lead to any great general insights on
fundamental principles of "neural computation" in the brain.
It is possible that all we will get from a good computer simulation is
good model that we can probe and study much more easily than a piece of
More information about the Comp-neuro